CS448f: Image Processing For Photography and Vision Blending and Pyramids
Blending We’ve aligned our images. What now? Averaging Weighted averaging min/max/median
Noise reduction by Averaging
2 Shots
4 Shots
8 Shots
16 Shots
Noise Reduction by Averaging We’re averaging random variables X and Y Both have variance S 2 Variance of X+Y = 2S 2 Std.Dev. of X+Y = sqrt(2). S Std.Dev of (X+Y)/2 = sqrt(2)/2. S Ie, every time we take twice as many photos, we reduce noise by sqrt(2)
Noise Reduction by Averaging Average 4 photos: noise gets reduced 2x Average 8 photos: noise gets reduced 3x Average 16 photos: noise gets reduced 4x
Noise Reduction by Median (demo)
Median v Average
Can we identify the bad pixels? They’re unlike their neighbours Instead of averaging, weighted average – where weight = similarity to neighbours
Weighted Average
Can we identify the bad pixels? They’re unlike their neighbours Instead of averaging, weighted average – where weight = similarity to neighbours Favors blurriness
Input
Other uses of Median Removing Transient Occluders (live demo) (Gates demo) (surf demo)
Panorama Stitching
Multiple Exposure Fusion
Focus Fusion
Pyramids We’ve been breaking images into two terms for a variety of apps – Coarse + Fine More generally we can break it into many terms: – Very coarse + finer + finer... + finest.
Pyramids We can do this by blurring more and more:
Pyramids And then (optionally) taking differences --
Pyramids The coarse layers can be stored at low res. Gaussian Pyramid Laplacian Pyramid
Pyramids How much memory does this use?
Pyramid Uses: Sampling arbitrarily sized Gaussians Equalizing an image – The different levels represent different frequency ranges – We can scale each frequency level and recombine Blending multiple images
Pyramid Blending Key Insight: – Coarse structure should blend very slowly between images (lots of feathering), while fine details should transition more quickly. More robust to tricky cases than plain old compositing
Inputs:
Compositing: Hard Mask
Compositing: Soft Mask
Multi-Band Blending
Exposure Fusion